The Effectiveness of Virtual Reality in Enhancing Students’ Reflective Abstraction in the Independent Curriculum

Various studies have explored the potential of VR in enhancing the quality of learning, particularly in fostering experiential and higher-order thinking. VR’s immersive and interactive nature aligns closely with Kolb’s experiential learning theory, which emphasizes four stages of knowledge transformation: concrete experience, reflective observation, abstract conceptualization, and active experimentation [14]. In this context, VR serves as a medium that not only provides authentic learning experiences but also facilitates reflective processing and conceptual understanding.

For example, Chang et al. [15] found that the use of VR significantly improved creative performance and conceptual thinking through immersive design tasks, supporting the notion that sensory-rich environments enhance learners' ability to reflect and abstract. Similarly, Dahlan et al. [16] highlighted the role of interactive visual content in improving learner engagement and conceptual grasp, although their study focused more on video learning than fully immersive VR. Both studies underscore the cognitive value of immersive media, but neither explicitly examines RA as a distinct construct, nor do they test mediating variables that explain how VR influences higher-order thinking outcomes.

In contrast, RA—originally proposed in Piagetian theory—refers to the learner’s ability to reconstruct and generalize knowledge through reflection on internalized experience. While various VR-based studies emphasize engagement or conceptual understanding, few assess RA directly or integrate it within a broader theoretical framework involving self-efficacy (SE) or learning engagement (LE). This represents a conceptual and methodological gap in the literature.

This study addresses that gap by synthesizing Kolb’s experiential cycle, Piaget’s RA framework, and Bandura’s concept of self-efficacy into a unified model. It contributes to the literature not only by investigating RA in a measurable way but also by examining the mediating effects of SE and LE using PLS-SEM. Situated within the Indonesian Independent Curriculum context—which emphasizes learner autonomy and reflective thinking—this study provides both theoretical advancement and empirical evidence on the mechanisms through which VR impacts student cognition.

4.1 Descriptive analysis of research data

Descriptive analysis is the initial stage in quantitative data processing that aims to describe the general characteristics of the data obtained through the research questionnaire. In this section, data from each main variable in the study were analyzed to determine the distribution of minimum, maximum, average (mean), and standard deviation values. These results provide an overview of participant responses to each variable and identify whether the data show sufficient diversity (variability) for further statistical analysis, including validity, reliability, or SEM. Thus, descriptive analysis serves as a basis for understanding the general pattern of responses and data quality before deeper interpretation. Table 2 presents the results of descriptive statistical analysis of the main variables in the study, namely VRLE, RA, SE, and LE, based on questionnaire data collected from respondents.

Table 2. Distribution of descriptive statistics

In Table 2, the analysis shows that the VRLE score has a minimum value of 7 and a maximum of 25, with an average of 19.02 and a standard deviation of 3.278. The RA variable has a score range of 8 to 25, with an average of 18.99 and a standard deviation of 3.433, indicating a relatively high but varied level of students' RA. The SE variable shows a minimum score of 8 and a maximum of 25, with an average of 19.04 and a standard deviation of 3.275, reflecting the level of students' confidence in completing tasks effectively. Meanwhile, the LE variable has a lower minimum value of 5 and a maximum of 25, with the highest average among other variables, namely 19.16, and a standard deviation of 3.449. These findings indicate that student engagement in VR-based learning is relatively high and consistent. The four variables have data distributions that tend to be normal and do not exhibit extreme deviations, making it suitable for subsequent analysis.

4.2 Linearity assumption test

Partial least squares (PLS) work on linear programs, and therefore any data scale, from nominal to ratio, can be analyzed in PLS as long as the data pattern is linear. This ensures that the hypothesis test in PLS can be estimated correctly. The linearity test aims to determine whether the relationship between the independent and dependent variables is linear. Since PLS does not include a built-in linearity test, SPSS was used to assess the linearity assumption. A relationship between the two variables is considered linear if the significance value of the test is smaller than the alpha value of 0.05 (5%). Table 3 presents the results of the linearity test.

Table 3. Distribution of descriptive statistics

Based on the results of the linearity assumption test presented in Table 3, all relationships between variables in the model have significance value of 0.000 (p < 0.05), which indicates that the relationship patterns are linear. Therefore, the SEM-PLS model is considered appropriate for use for further analysis.

4.3 Outer model

The outer model is a model with calculation results based on calculations using the PLS program. Confirmatory factor analysis was utilized to determine whether the existing indicators accurately explain the constructs [19]. The purpose of the measurement model is to describe how well the indicators in this study can be used as measurement instruments for latent variables [20].

Validity evaluation of the measurement model was performed by examining the estimated factor loadings. A variable demonstrates good validity for its construct or latent variable if the t-value of its factor loading exceeds the critical value (≥ 1.96) and/or its standard factor loading is ≥ 0.50 [21]. Reliability evaluation in PLS used composite reliability (CR) ≥ 0.70 and average variance extracted (AVE) ≥ 0.50 [22]. Table 4 shows the results of the validity and reliability evaluation.

Table 4. Recapitulation of validity and reliability evaluation results

Based on Table 4, all reflective indicators have loading factor values ≥ 0.50 and AVE values ≥ 0.50, confirming that all indicators are valid. The reliability calculation results also show that the CR values are ≥ 0.70, indicating that the constructs are reliable. Therefore, all latent variables have adequate and appropriate indicators. To determine which indicators most strongly contribute to each latent construct, the dominant indicators are as follows:

1. The best indicator in representing the VRLE variable is Vrle.4 (“VR technology helps me understand abstract or complex concepts”) with the highest factor loading of 0.802. To improve the VRLE variable, Vrle.4 should be prioritized.
2. The best indicator in representing the RA variable is RA.1 (“I am able to draw conclusions from the learning experiences I have had”) with the highest factor loading of 0.842. To improve the RA variable, RA.1 should be prioritized.
3. The best indicator in representing the SE variable is SE.1 (“I am confident that I can complete the learning tasks given with the help of VR technology”) with the highest factor loading of 0.807. To improve the SE variable, SE.1 should be prioritized.
4. The best indicator in representing the LE variable is LE.3 (“I feel that learning with VR makes me more motivated”) with the highest factor loading of 0.815. To improve the LE variable, LE.3 should be prioritized.

4.4 HTMT test (discriminant validity)

The HTMT test is an approach used to assess discriminant validity in structural equation modeling–partial least squares (SEM-PLS) [23]. Discriminant validity indicates the extent to which a construct in the model is truly different (discrete) from other constructs [24]. This means that each latent variable must be able to explain a unique concept and not overlap with other constructs. The HTMT is calculated based on the average ratio of correlation between indicators from different constructs (heterotrait) to the correlation between indicators in the same construct (monotrait). The resulting HTMT value is compared with a certain threshold value. According to Ab Hamid et al. [25], a model demonstrates good discriminant validity if the HTMT value is below 0.90. Values exceeding this threshold indicate a problem in distinguishing between constructs, potentially compromising the validity of the model. Table 5 shows the results of the HTMT test.

The results of the discriminant validity test using the HTMT method indicate that all HTMT values between constructs are below the 0.90 threshold, which is the general criterion for good discriminant validity [26]. The HTMT values between LE and other constructs ranges from 0.701 to 0.845, with the highest value of 0.845 observed between LE and RA. The relationship between SE and RA is 0.808, while the relationship between SE and LE is 0.829. The lowest value is for the relationship between VRLE and LE, which is 0.701.

These values indicate that each construct (LE, RA, SE, and VRLE) has clear conceptual differences without significant overlap, confirming that the instrument used in this study has met the requirements for discriminant validity.

Table 5. HTMT test results

4.5 Hypothesis testing (path analysis)

This section evaluates the coefficients that indicate causal relationships or influences among latent variables. A causal relationship is considered statistically insignificant if the t-statistic falls between -1.96 and 1.96 at a significance level of 0.05. The calculation of these coefficients is presented in Table 6.

Table 6. SEM-PLS path analysis results

The results of the path analysis show that all relationships between variables in the model are positive and statistically significant. The VRLE has a significant effect on SE (β = 0.557; t = 8.191), LE (β = 0.271; t = 2.889), and RA (β = 0.263; t = 2.281). SE also has a significant effect on LE (β = 0.492; t = 4.941) and RA (β = 0.269; t = 2.272). Meanwhile, LE makes a significant contribution to RA (β = 0.349; t = 2.743). Since all t-values exceed the critical threshold of 1.96, all proposed hypotheses are accepted, indicating the structural validity of the model. The path coefficients in the structural model and the weight values of the manifest variable factors in the measurement model are illustrated in Figure 2.

2.png

Figure 2. Path diagram of measurement model and structural model (overall)

Based on the path diagram in Figure 2, the RA variable is more dominantly influenced by the LE variable, with the highest path coefficient of 0.349. The LE variable, in turn, is more dominantly influenced by the SE variable, with a path coefficient of 0.492, and SE is influenced by VRLE, with a path coefficient of 0.557. The most dominant indicator representing VRLE is Vrle.4 (“VR technology helps me understand abstract or complex concepts”), with the highest factor loading of 0.802. Therefore, if policymakers aim to improve the RA variable, they should focus on improving Vrle.4 as the key performance indicator, addressing it as a strategic priority for policy evaluation.

4.6 Analysis of mediating variables (indirect effect)

The analysis of mediating variables can be conducted through two approaches: coefficient differences and coefficient multiplication. The coefficient difference approach uses an examination method by conducting analysis with and without involving mediating variables, while the multiplication method is carried out using the Sobel method. In this study, the detection was performed using the coefficient multiplication approach and the Sobel test [27]. The results of this analysis are presented in Table 7.

Table 7. Indirect effects among latent variables

Based on Table 7, the mediation analysis reveals several significant indirect effects between VR-based learning experience (VRLE) and reflective abstraction (RA) through the mediating variables self-efficacy (SE) and learning engagement (LE).

The indirect effect of VRLE → SE → RA is 0.194 (t = 2.367, p < 0.05). The indirect effect of VRLE → LE → RA is 0.095 (t = 2.367, p < 0.05). The indirect effect of SE → LE → RA is 0.172 (t = 2.393, p < 0.05). The indirect effect of VRLE → SE → LE → RA is 0.096 (t = 2.006, p < 0.05). The total indirect effect of VRLE on RA is 0.385, based on the sum of the three mediated pathways: 0.194 (SE)+ 0.095 (LE)+ 0.096 (SE→LE)=0.385.

Assuming the total effect of VRLE on RA is approximately 0.50, the proportion of mediation for each path was calculated using the following formula. The resulting mediation proportions are as follows:

$V R L E \rightarrow S E \rightarrow R A: \frac{0.194}{0.50} \times 100 \%=38.5 \%$

$V R L E \rightarrow L E \rightarrow R A: \frac{0.095}{0.50} \times 100 \%=19.0 \%$

$V R L E \rightarrow S E \rightarrow L E \rightarrow R A: \frac{0.096}{0.50} \times 100 \%=19.2 \%$

These results suggest that SE is the dominant mediator, accounting for nearly twice the influence of LE alone. Furthermore, the sequential path through SE and LE contributes an additional 19.2%, highlighting the combined motivational and behavioral mechanism by which VRLE enhances students' reflective abstraction.

While the analysis supports a predictive relationship between VRLE and RA via SE and LE, the cross-sectional design limits direct causal inference. Nonetheless, the findings are consistent with causal hypotheses proposed by experiential learning and self-efficacy theories.